@grida/ai-models 0.0.0-stage → 1.20261011.2
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/LICENSE +21 -0
- package/PROVENANCE.md +681 -0
- package/README.md +39 -2
- package/labs.json +178 -0
- package/logos/alibaba.svg +1 -0
- package/logos/anthropic.svg +4 -0
- package/logos/black-forest-labs.svg +1 -0
- package/logos/bria.png +0 -0
- package/logos/bytedance.png +0 -0
- package/logos/elevenlabs.svg +4 -0
- package/logos/google.png +0 -0
- package/logos/meta.svg +72 -0
- package/logos/microsoft.svg +1 -0
- package/logos/openai.svg +3 -0
- package/logos/recraft-ai.svg +4 -0
- package/logos/tencent.png +0 -0
- package/logos/tripo.svg +13 -0
- package/logos/xai.svg +1 -0
- package/manifest.json +95 -0
- package/models.json +2802 -0
- package/package.json +37 -4
- package/schemas/labs.schema.json +68 -0
- package/schemas/models.schema.json +230 -0
package/LICENSE
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
1
|
+
MIT License
|
|
2
|
+
|
|
3
|
+
Copyright (c) 2026 Grida
|
|
4
|
+
|
|
5
|
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
6
|
+
of this software and associated documentation files (the "Software"), to deal
|
|
7
|
+
in the Software without restriction, including without limitation the rights
|
|
8
|
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
9
|
+
copies of the Software, and to permit persons to whom the Software is
|
|
10
|
+
furnished to do so, subject to the following conditions:
|
|
11
|
+
|
|
12
|
+
The above copyright notice and this permission notice shall be included in all
|
|
13
|
+
copies or substantial portions of the Software.
|
|
14
|
+
|
|
15
|
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
16
|
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
17
|
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
18
|
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
19
|
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
20
|
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
21
|
+
SOFTWARE.
|
package/PROVENANCE.md
ADDED
|
@@ -0,0 +1,681 @@
|
|
|
1
|
+
# Catalogue research notes
|
|
2
|
+
|
|
3
|
+
Source-specific qualifications and references for the values in
|
|
4
|
+
[`models.json`](./models.json). The JSON holds the values; these notes explain
|
|
5
|
+
where they came from and any caveat a value cannot carry itself. Update a note
|
|
6
|
+
in the same change as the fact it qualifies.
|
|
7
|
+
|
|
8
|
+
Each `##` heading names the entry or declaration a note qualifies. The headings
|
|
9
|
+
are code fragments carried over from the TypeScript literals the catalogue was
|
|
10
|
+
first written as in [gridaco/grida](https://github.com/gridaco/grida), so they
|
|
11
|
+
keep that historical spelling (for example `imageInputMimes` rather than
|
|
12
|
+
`image_input_mimes`).
|
|
13
|
+
|
|
14
|
+
## `const OPENAI_IMAGE_INPUT_MIMES = [`
|
|
15
|
+
|
|
16
|
+
Provider-family capabilities are kept private so catalogue entries remain
|
|
17
|
+
explicit while sharing one source-backed value. Do not derive these from
|
|
18
|
+
`multimodal`; a future model may be multimodal without a documented image
|
|
19
|
+
input format set.
|
|
20
|
+
https://developers.openai.com/api/docs/guides/images-vision#image-input-requirements
|
|
21
|
+
|
|
22
|
+
## `const ANTHROPIC_IMAGE_INPUT_MIMES = [`
|
|
23
|
+
|
|
24
|
+
https://platform.claude.com/docs/en/build-with-claude/vision#supported-formats
|
|
25
|
+
|
|
26
|
+
## `const GOOGLE_IMAGE_INPUT_MIMES = [`
|
|
27
|
+
|
|
28
|
+
https://ai.google.dev/gemini-api/docs/image-understanding#supported-image-formats
|
|
29
|
+
|
|
30
|
+
## `const OPENAI_LONG_CONTEXT_PRICING = {`
|
|
31
|
+
|
|
32
|
+
OpenAI bills the full request at these multipliers once its total input
|
|
33
|
+
exceeds 272K tokens. The same rule is published for GPT-5.5, every
|
|
34
|
+
GPT-5.6 family member, and GPT-6 Astra.
|
|
35
|
+
https://developers.openai.com/api/docs/models/gpt-5.5
|
|
36
|
+
https://developers.openai.com/api/docs/models/gpt-5.6-sol
|
|
37
|
+
https://developers.openai.com/api/docs/models/gpt-6-astra
|
|
38
|
+
|
|
39
|
+
## `"openai/gpt-5.6-sol": {`
|
|
40
|
+
|
|
41
|
+
Base rates; OPENAI_LONG_CONTEXT_PRICING represents the request-wide
|
|
42
|
+
band that applies above 272K total input tokens.
|
|
43
|
+
|
|
44
|
+
Sol is cheaper than GPT-5.5, the card directly above it. It was
|
|
45
|
+
introduced carrying 5.5's rates verbatim; these are OpenAI's own.
|
|
46
|
+
https://developers.openai.com/api/docs/models/gpt-5.6-sol
|
|
47
|
+
|
|
48
|
+
## `"openai/gpt-6-astra": {`
|
|
49
|
+
|
|
50
|
+
OpenAI's September 3 introduction remained a limited rollout. The
|
|
51
|
+
September 4 date below is the exact Vercel AI Gateway route's broad availability,
|
|
52
|
+
which is the release fact relevant to this route card.
|
|
53
|
+
https://openai.com/products/release-notes/
|
|
54
|
+
https://vercel.com/ai-gateway/models/gpt-6-astra
|
|
55
|
+
|
|
56
|
+
## `"anthropic/claude-opus-5.5": {`
|
|
57
|
+
|
|
58
|
+
Anthropic API ID `claude-opus-5-5`; this is the Vercel AI Gateway ID, whose
|
|
59
|
+
model feed lists the same rates. Cache reads are 0.05x base input ($0.20);
|
|
60
|
+
`cache_write` is the 5-minute write ($5; the 1-hour write is $8). Released
|
|
61
|
+
2026-09-22 as generally available; only its fast mode is a research preview.
|
|
62
|
+
https://platform.claude.com/docs/en/models/opus-5-5/overview
|
|
63
|
+
https://platform.claude.com/docs/en/about-claude/pricing
|
|
64
|
+
https://vercel.com/ai-gateway/models/claude-opus-5.5
|
|
65
|
+
|
|
66
|
+
Claude Opus 5 stays catalogued. Opus 5.5 is cheaper, but it is not a drop-in
|
|
67
|
+
successor: thinking can't be disabled, forced tool use (`tool_choice`
|
|
68
|
+
`any`/`tool`) returns an error, and `computer_20251124` is not accepted on
|
|
69
|
+
the Claude API and Google Cloud. Opus 5 is still the choice for those.
|
|
70
|
+
https://platform.claude.com/docs/en/models/opus-5-5/whats-new-opus-5-5
|
|
71
|
+
|
|
72
|
+
## `"anthropic/claude-haiku-5.5": {`
|
|
73
|
+
|
|
74
|
+
Anthropic API ID `claude-haiku-5-5`; this is the Vercel AI Gateway ID. Priced
|
|
75
|
+
by prompt length: a request whose prompt is over 100,000 tokens, counting all
|
|
76
|
+
input tokens including cache reads and writes, pays $0.50 input, $2.50
|
|
77
|
+
output, $0.05 cache read and $0.625 5-minute cache write, 5x each base rate.
|
|
78
|
+
`long_context` records that band; its input multiplier also applies to cache
|
|
79
|
+
reads and writes. The gateway's model feed publishes the same tiers. The
|
|
80
|
+
1.1x US-only inference multiplier (`inference_geo`) also applies to the
|
|
81
|
+
higher band.
|
|
82
|
+
https://platform.claude.com/docs/en/models/haiku-5-5/overview
|
|
83
|
+
https://platform.claude.com/docs/en/about-claude/pricing
|
|
84
|
+
https://vercel.com/ai-gateway/models/claude-haiku-5.5
|
|
85
|
+
|
|
86
|
+
## `"anthropic/claude-sonnet-5": {`
|
|
87
|
+
|
|
88
|
+
$2/$10 is the standard rate, not a live discount: it launched as an
|
|
89
|
+
introductory rate and Anthropic made it permanent, cancelling the
|
|
90
|
+
announced step up to $3/$15. Do not restore the higher card.
|
|
91
|
+
https://platform.claude.com/docs/en/about-claude/pricing
|
|
92
|
+
|
|
93
|
+
## `"anthropic/claude-fable-5.1": {`
|
|
94
|
+
|
|
95
|
+
Same input/output/cacheWrite card as Claude Fable 5, with cache reads
|
|
96
|
+
cut to $0.25/MTok. Not a drop-in successor — forced tool choice
|
|
97
|
+
(`tool_choice` `any`/`tool`) is rejected here — which is why Fable 5
|
|
98
|
+
stays catalogued rather than being removed.
|
|
99
|
+
https://platform.claude.com/docs/en/models/fable-5-1/overview
|
|
100
|
+
|
|
101
|
+
## `"anthropic/claude-fable-5": {`
|
|
102
|
+
|
|
103
|
+
Unlike Fable 5.1, this version accepts forced tool choice.
|
|
104
|
+
|
|
105
|
+
## `"anthropic/claude-opus-5": {`
|
|
106
|
+
|
|
107
|
+
Drop-in successor to Opus 4.8 at the same rate card.
|
|
108
|
+
|
|
109
|
+
## `"google/gemini-3.8-flash": {`
|
|
110
|
+
|
|
111
|
+
Google's cache model is read + hourly storage (no one-time write
|
|
112
|
+
premium that matches `cacheWrite` semantics), so the field is omitted.
|
|
113
|
+
|
|
114
|
+
Steady-state rates. Google is promoting $0.75 in / $3.75 out / $0.075
|
|
115
|
+
cacheRead through 2026-12-31; these are the prices that apply from
|
|
116
|
+
2027-01-01. Checking Google's page before then will show the lower
|
|
117
|
+
set — that is the promotion, not a correction. Gemini 3.8 Flash is GA,
|
|
118
|
+
and this exact id is live on both Vercel AI Gateway and OpenRouter.
|
|
119
|
+
https://ai.google.dev/gemini-api/docs/models/gemini-3.8-flash
|
|
120
|
+
https://ai.google.dev/gemini-api/docs/pricing
|
|
121
|
+
https://vercel.com/ai-gateway/models/gemini-3.8-flash
|
|
122
|
+
https://openrouter.ai/google/gemini-3.8-flash
|
|
123
|
+
|
|
124
|
+
## `"google/gemini-3.7-flash": {`
|
|
125
|
+
|
|
126
|
+
3.8 improves accuracy and reliability at the same rate, but Google
|
|
127
|
+
still recommends 3.7 when compute efficiency matters because 3.8 can
|
|
128
|
+
consume more tokens.
|
|
129
|
+
https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/guides/gemini-3-8-flash
|
|
130
|
+
|
|
131
|
+
## `longContext: {`
|
|
132
|
+
|
|
133
|
+
Google bills the full request at $4 in / $18 out / $0.40
|
|
134
|
+
cacheRead above 200K tokens. `inputMultiplier` covers every
|
|
135
|
+
input bucket, so cacheRead 0.2 x 2 lands on $0.40 without a
|
|
136
|
+
separate field.
|
|
137
|
+
https://ai.google.dev/gemini-api/docs/pricing
|
|
138
|
+
|
|
139
|
+
## `const GPT_IMAGE_2_5_FAL_PRICING: PerTokenPricing = {`
|
|
140
|
+
|
|
141
|
+
Both 2.5 variants use this fal meter. Keep the separate text-output
|
|
142
|
+
rate: fal publishes it even though its result schema exposes only images.
|
|
143
|
+
https://fal.ai/models/openai/gpt-image-2.5/flare/text-to-image
|
|
144
|
+
https://fal.ai/models/openai/gpt-image-2.5/sunburst/text-to-image
|
|
145
|
+
|
|
146
|
+
## `const GPT_IMAGE_2_5_VERCEL_AI_GATEWAY_PRICING: PerTokenPricing = {`
|
|
147
|
+
|
|
148
|
+
Each serving provider publishes its own meter; do not copy fal's extra
|
|
149
|
+
image-cache/text-output rates into providers that do not advertise them.
|
|
150
|
+
https://ai-gateway.vercel.sh/v1/models (verified 2026-09-09 KST)
|
|
151
|
+
|
|
152
|
+
## `const GPT_IMAGE_2_5_OPENROUTER_PRICING: PerTokenPricing = {`
|
|
153
|
+
|
|
154
|
+
Both variants' /api/v1/images/models/{id}/endpoints publish this meter.
|
|
155
|
+
https://openrouter.ai/api/v1/images/models/openai/gpt-image-2.5-flare/endpoints
|
|
156
|
+
https://openrouter.ai/api/v1/images/models/openai/gpt-image-2.5-sunburst/endpoints
|
|
157
|
+
|
|
158
|
+
## `"openai/gpt-image-2": {`
|
|
159
|
+
|
|
160
|
+
---
|
|
161
|
+
|
|
162
|
+
OpenAI
|
|
163
|
+
-----------------------------------------------------------------
|
|
164
|
+
|
|
165
|
+
https://developers.openai.com/api/docs/models/gpt-image-2
|
|
166
|
+
|
|
167
|
+
## `transparent_background: true,`
|
|
168
|
+
|
|
169
|
+
Native transparency added 2026-08-20; provider exposure differs.
|
|
170
|
+
https://developers.openai.com/api/docs/changelog
|
|
171
|
+
|
|
172
|
+
## `providers: {`
|
|
173
|
+
|
|
174
|
+
ids/prices verified 2026-06-29, see github.com/gridaco/grida/issues/908
|
|
175
|
+
|
|
176
|
+
## `transparent_background: null,`
|
|
177
|
+
|
|
178
|
+
This provider's page does not establish background support.
|
|
179
|
+
https://vercel.com/ai-gateway/models/gpt-image-2 (2026-09-09)
|
|
180
|
+
|
|
181
|
+
## `transparent_background: false,`
|
|
182
|
+
|
|
183
|
+
Published background enum is auto | opaque (2026-09-09).
|
|
184
|
+
https://openrouter.ai/api/v1/images/models/openai/gpt-image-2/endpoints
|
|
185
|
+
|
|
186
|
+
## `references: { id: "openai/gpt-image-2", max: 16 },`
|
|
187
|
+
|
|
188
|
+
input_references advertised by OpenRouter (0–16), 2026-07-01.
|
|
189
|
+
|
|
190
|
+
## `},`
|
|
191
|
+
|
|
192
|
+
Inherits native transparency: fal-ai/gpt-image-2's published
|
|
193
|
+
background enum includes transparent (verified 2026-09-09).
|
|
194
|
+
|
|
195
|
+
## `"openai/gpt-image-2.5-flare": {`
|
|
196
|
+
|
|
197
|
+
Released as two distinct models. All three providers now list both
|
|
198
|
+
variants; fal separates generation and edit endpoint ids, while
|
|
199
|
+
OpenRouter advertises input_references on the same id (2026-09-09 KST).
|
|
200
|
+
|
|
201
|
+
## `transparent_background: true,`
|
|
202
|
+
|
|
203
|
+
Both fal generation and edit schemas expose background=transparent.
|
|
204
|
+
https://fal.ai/models/openai/gpt-image-2.5/flare/text-to-image/api
|
|
205
|
+
https://fal.ai/models/openai/gpt-image-2.5/flare/edit/api
|
|
206
|
+
|
|
207
|
+
## `url: "https://vercel.com/ai-gateway/models/gpt-image-2.5-flare",`
|
|
208
|
+
|
|
209
|
+
The model page exposes background=transparent; reference input
|
|
210
|
+
is not established by the Vercel AI Gateway feed (2026-09-09 KST).
|
|
211
|
+
|
|
212
|
+
## `transparent_background: false,`
|
|
213
|
+
|
|
214
|
+
Endpoint schema's background enum is auto | opaque.
|
|
215
|
+
|
|
216
|
+
## `constraints: {`
|
|
217
|
+
|
|
218
|
+
Model-specific fal documentation overrides its generic ImageSize
|
|
219
|
+
component's wider bounds. The API also accepts auto-sized output.
|
|
220
|
+
https://fal.ai/models/openai/gpt-image-2.5/flare/text-to-image/api
|
|
221
|
+
|
|
222
|
+
## `avg_cost_usd: 0.055,`
|
|
223
|
+
|
|
224
|
+
High 1024² output estimate is $0.05268, plus a small input allowance.
|
|
225
|
+
Not a fixed per-image price: actual cost depends on all billed tokens.
|
|
226
|
+
https://developers.openai.com/api/docs/guides/image-generation
|
|
227
|
+
|
|
228
|
+
## `transparent_background: true,`
|
|
229
|
+
|
|
230
|
+
Both fal generation and edit schemas expose background=transparent.
|
|
231
|
+
https://fal.ai/models/openai/gpt-image-2.5/sunburst/text-to-image/api
|
|
232
|
+
https://fal.ai/models/openai/gpt-image-2.5/sunburst/edit/api
|
|
233
|
+
|
|
234
|
+
## `url: "https://vercel.com/ai-gateway/models/gpt-image-2.5-sunburst",`
|
|
235
|
+
|
|
236
|
+
The model page exposes background=transparent; reference input
|
|
237
|
+
is not established by the Vercel AI Gateway feed (2026-09-09 KST).
|
|
238
|
+
|
|
239
|
+
## `transparent_background: false,`
|
|
240
|
+
|
|
241
|
+
Endpoint schema's background enum is auto | opaque.
|
|
242
|
+
|
|
243
|
+
## `constraints: {`
|
|
244
|
+
|
|
245
|
+
https://fal.ai/models/openai/gpt-image-2.5/sunburst/text-to-image/api
|
|
246
|
+
|
|
247
|
+
## `"openai/gpt-image-1.5": {`
|
|
248
|
+
|
|
249
|
+
https://developers.openai.com/api/docs/models/gpt-image-1.5
|
|
250
|
+
|
|
251
|
+
## `constraints: null,`
|
|
252
|
+
|
|
253
|
+
Preset-only — provider rejects arbitrary sizes.
|
|
254
|
+
|
|
255
|
+
## `pricing: {`
|
|
256
|
+
|
|
257
|
+
https://developers.openai.com/api/docs/models/gpt-image-1.5
|
|
258
|
+
|
|
259
|
+
## `"openai/gpt-image-1-mini": {`
|
|
260
|
+
|
|
261
|
+
https://developers.openai.com/api/docs/models/gpt-image-1-mini
|
|
262
|
+
|
|
263
|
+
## `constraints: null,`
|
|
264
|
+
|
|
265
|
+
Preset-only — provider rejects arbitrary sizes.
|
|
266
|
+
|
|
267
|
+
## `pricing: {`
|
|
268
|
+
|
|
269
|
+
https://developers.openai.com/api/docs/models/gpt-image-1-mini
|
|
270
|
+
|
|
271
|
+
## `"google/gemini-3.1-flash-image-preview": {`
|
|
272
|
+
|
|
273
|
+
---
|
|
274
|
+
|
|
275
|
+
Google (multimodal LLMs with native image output)
|
|
276
|
+
-----------------------------------------------------------------
|
|
277
|
+
|
|
278
|
+
python .tools/model_info.py --image gemini-3.1-flash-image
|
|
279
|
+
Vercel AI Gateway pricing: $0.50/MTok input, $3.00/MTok output
|
|
280
|
+
|
|
281
|
+
## `providers: {`
|
|
282
|
+
|
|
283
|
+
"Nano Banana 2"; ids/prices verified 2026-06-29, see issues/908
|
|
284
|
+
|
|
285
|
+
## `vercel: {`
|
|
286
|
+
|
|
287
|
+
Vercel AI Gateway serves the graduated `google/gemini-3.1-flash-image`
|
|
288
|
+
and the `-preview` alias at identical rates (feed, 2026-09-02).
|
|
289
|
+
Bindings call the graduated id; the canonical key above stays
|
|
290
|
+
`-preview` because it is persisted in selections and published
|
|
291
|
+
in the catalogue snapshot — renaming it is a separate change.
|
|
292
|
+
|
|
293
|
+
## `references: { id: "google/gemini-3.1-flash-image", max: 14 },`
|
|
294
|
+
|
|
295
|
+
input_references advertised by OpenRouter (0–14), 2026-07-01.
|
|
296
|
+
|
|
297
|
+
## `fal: {`
|
|
298
|
+
|
|
299
|
+
fal's graduated endpoint is `fal-ai/nano-banana-2`; same $0.08
|
|
300
|
+
per 1K image as the `-preview` endpoint (2K ×1.5, 4K ×2).
|
|
301
|
+
|
|
302
|
+
## `"google/gemini-3-pro-image": {`
|
|
303
|
+
|
|
304
|
+
python .tools/model_info.py --image gemini-3-pro-image
|
|
305
|
+
Vercel AI Gateway pricing: $2.00/MTok input, $12.00/MTok output
|
|
306
|
+
|
|
307
|
+
## `providers: {`
|
|
308
|
+
|
|
309
|
+
"Nano Banana Pro"; ids/prices verified 2026-06-29, see issues/908
|
|
310
|
+
|
|
311
|
+
## `references: { id: "google/gemini-3-pro-image-preview", max: 14 },`
|
|
312
|
+
|
|
313
|
+
input_references advertised by OpenRouter (0–14), 2026-07-01.
|
|
314
|
+
|
|
315
|
+
## `"google/gemini-3.1-flash-lite-image": {`
|
|
316
|
+
|
|
317
|
+
"Nano Banana 2 Lite" — GA 2026-06-30. The cost/speed tier of the 3.1
|
|
318
|
+
Flash family: ~half of Nano Banana 2's meter, and 1K-only output
|
|
319
|
+
(2K/4K unsupported — the differentiator). Vercel AI Gateway + OpenRouter both
|
|
320
|
+
meter it at $0.25/$1.50 (verified 2026-07-01); fal id not verified,
|
|
321
|
+
so left out. OpenRouter doesn't advertise input_references for the
|
|
322
|
+
Lite (t2i only per its model page), so no `references` (TOOL-DESIGN:
|
|
323
|
+
no unverified capability).
|
|
324
|
+
|
|
325
|
+
## `avg_cost_usd: 0.034,`
|
|
326
|
+
|
|
327
|
+
Published 1K per-image cost (the card default): $0.034 = 1120 tokens
|
|
328
|
+
× $30/1M image-output (Google/Vercel AI Gateway changelog, 2026-07-01). The
|
|
329
|
+
budget meter charges this per image, so it must be the real cost.
|
|
330
|
+
|
|
331
|
+
## `"bfl/flux-2-pro": {`
|
|
332
|
+
|
|
333
|
+
---
|
|
334
|
+
|
|
335
|
+
Black Forest Labs (via Vercel AI Gateway)
|
|
336
|
+
-----------------------------------------------------------------
|
|
337
|
+
|
|
338
|
+
https://vercel.com/docs/ai-gateway/capabilities/image-generation/ai-sdk
|
|
339
|
+
https://docs.bfl.ml/pricing
|
|
340
|
+
|
|
341
|
+
## `providers: {`
|
|
342
|
+
|
|
343
|
+
All three providers meter $0.03 per megapixel (Vercel AI Gateway model page,
|
|
344
|
+
OpenRouter endpoint `cost_usd`/megapixel, fal "first megapixel");
|
|
345
|
+
represented as flat at the 1MP baseline. The Vercel AI Gateway binding shipped
|
|
346
|
+
at $0.06 — a 2x hosted over-billing — corrected 2026-09-02. The
|
|
347
|
+
Vercel AI Gateway feed carries no `pricing` for BFL cards, so the page is the
|
|
348
|
+
source.
|
|
349
|
+
|
|
350
|
+
## `references: { id: "black-forest-labs/flux.2-pro", max: 8 },`
|
|
351
|
+
|
|
352
|
+
input_references advertised by OpenRouter (0–8), 2026-07-01.
|
|
353
|
+
|
|
354
|
+
## `"bfl/flux-2-max": {`
|
|
355
|
+
|
|
356
|
+
---
|
|
357
|
+
|
|
358
|
+
Black Forest Labs — Flux 2 Max
|
|
359
|
+
-----------------------------------------------------------------
|
|
360
|
+
|
|
361
|
+
BFL's top Flux 2 line (2025-12-16). $0.07 per megapixel on all three
|
|
362
|
+
(Vercel AI Gateway model page — the feed carries no BFL pricing; OpenRouter
|
|
363
|
+
`cost_usd`/megapixel; fal "first megapixel", +$0.03 each additional).
|
|
364
|
+
Represented as flat at the 1MP baseline. Verified 2026-09-02.
|
|
365
|
+
|
|
366
|
+
## `references: { id: "black-forest-labs/flux.2-max", max: 8 },`
|
|
367
|
+
|
|
368
|
+
OpenRouter `supported_parameters.input_references` 0–8
|
|
369
|
+
(2026-09-02). As on Flux 2 Pro.
|
|
370
|
+
|
|
371
|
+
## `constraints: { min_edge: 256, max_edge: 1440 },`
|
|
372
|
+
|
|
373
|
+
Same envelope as Flux 2 Pro pending a vendor spec for Max.
|
|
374
|
+
|
|
375
|
+
## `providers: {`
|
|
376
|
+
|
|
377
|
+
$0.04 on both providers: Vercel AI Gateway feed `pricing.image` + fal's model
|
|
378
|
+
page ("Fixed $0.04 cost per image edit"), verified 2026-09-02.
|
|
379
|
+
|
|
380
|
+
## `"bytedance/seedream-5.0-pro": {`
|
|
381
|
+
|
|
382
|
+
---
|
|
383
|
+
|
|
384
|
+
ByteDance — Seedream 5.0 Pro
|
|
385
|
+
-----------------------------------------------------------------
|
|
386
|
+
|
|
387
|
+
Vercel AI Gateway feed `pricing.image` $0.035 (the model page's rate table shows
|
|
388
|
+
$0.04 while its copy says $0.035 — the feed is the billing contract);
|
|
389
|
+
OpenRouter `cost_usd` $0.045 at 1K ($0.09 high-res, +$0.003 per
|
|
390
|
+
input image, not modelled); fal $0.0675 for ≤1536² area, $0.135 up
|
|
391
|
+
to 2048². Represented at the 1024² baseline. Verified 2026-09-02.
|
|
392
|
+
|
|
393
|
+
## `references: { id: "bytedance-seed/seedream-5-0-pro", max: 14 },`
|
|
394
|
+
|
|
395
|
+
OpenRouter `supported_parameters.input_references` 0–14
|
|
396
|
+
(2026-09-02). Same endpoint as t2i, as on 4.5.
|
|
397
|
+
|
|
398
|
+
## `constraints: {`
|
|
399
|
+
|
|
400
|
+
fal: total pixels between 1024x1024 and 2048x2048, aspect ratio
|
|
401
|
+
between 1:16 and 16:1 — an area bound, not a per-edge one, so a
|
|
402
|
+
512x2048 request is in-envelope and a 2048x2048 one is the ceiling.
|
|
403
|
+
|
|
404
|
+
## `"bytedance/seedream-5.0-lite": {`
|
|
405
|
+
|
|
406
|
+
---
|
|
407
|
+
|
|
408
|
+
ByteDance — Seedream 5.0 Lite
|
|
409
|
+
-----------------------------------------------------------------
|
|
410
|
+
|
|
411
|
+
Universal: $0.035/img on all three (Vercel AI Gateway feed `pricing.image`,
|
|
412
|
+
OpenRouter `cost_usd`, fal page payload), verified 2026-09-02. A
|
|
413
|
+
2K–4K model: fal scales requests below 2560x1440 up to its floor.
|
|
414
|
+
|
|
415
|
+
## `references: { id: "bytedance-seed/seedream-5-0-lite", max: 14 },`
|
|
416
|
+
|
|
417
|
+
OpenRouter `supported_parameters.input_references` 0–14
|
|
418
|
+
(2026-09-02).
|
|
419
|
+
|
|
420
|
+
## `constraints: { min_pixels: 3_686_400, max_pixels: 16_777_216 },`
|
|
421
|
+
|
|
422
|
+
fal: total pixels between 2560x1440 and 4096x4096 (requests below
|
|
423
|
+
the floor are scaled up to it); Vercel AI Gateway and OpenRouter serve 2K/4K
|
|
424
|
+
only. An area envelope, so the default is a 2K request.
|
|
425
|
+
|
|
426
|
+
## `"bytedance/seedream-4.5": {`
|
|
427
|
+
|
|
428
|
+
---
|
|
429
|
+
|
|
430
|
+
ByteDance — Seedream 4.5
|
|
431
|
+
-----------------------------------------------------------------
|
|
432
|
+
|
|
433
|
+
$0.04/img on all three providers (re-verified 2026-09-02).
|
|
434
|
+
|
|
435
|
+
## `references: { id: "bytedance-seed/seedream-4.5", max: 14 },`
|
|
436
|
+
|
|
437
|
+
i2i verified live 2026-07-01 (OpenRouter /api/v1/images
|
|
438
|
+
input_references; same endpoint as t2i). max from
|
|
439
|
+
supported_parameters.input_references.
|
|
440
|
+
|
|
441
|
+
## `"xai/grok-imagine-image-2.0": {`
|
|
442
|
+
|
|
443
|
+
---
|
|
444
|
+
|
|
445
|
+
SpaceXAI — Grok Imagine Image 2.0
|
|
446
|
+
-----------------------------------------------------------------
|
|
447
|
+
|
|
448
|
+
Tiered by quality (low/medium) × resolution (1K/2K); the same four
|
|
449
|
+
rates on all three providers (Vercel AI Gateway feed
|
|
450
|
+
`image_dimension_quality_pricing`, OpenRouter `cost_usd` variants,
|
|
451
|
+
fal page). OpenRouter also bills $0.01 per input image (not
|
|
452
|
+
modelled). Verified 2026-09-02.
|
|
453
|
+
|
|
454
|
+
## `references: { id: "x-ai/grok-imagine-image-2.0", max: 3 },`
|
|
455
|
+
|
|
456
|
+
OpenRouter `supported_parameters.input_references` 0–3
|
|
457
|
+
(2026-09-02). OR bills $0.01 per input image on top (not modelled).
|
|
458
|
+
|
|
459
|
+
## `"meta/muse-image-1.0": {`
|
|
460
|
+
|
|
461
|
+
---
|
|
462
|
+
|
|
463
|
+
Meta — Muse Image 1.0
|
|
464
|
+
-----------------------------------------------------------------
|
|
465
|
+
|
|
466
|
+
Meta's agentic image model (2026-08-26): $0.01/img on Vercel AI Gateway (feed
|
|
467
|
+
`pricing.image`) and fal (page payload). OpenRouter lists it but
|
|
468
|
+
exposes no serving endpoint.
|
|
469
|
+
fal exposes aspect ratio only (no size control). Verified 2026-09-02.
|
|
470
|
+
|
|
471
|
+
## `constraints: null,`
|
|
472
|
+
|
|
473
|
+
Muse picks output dimensions from the aspect ratio; neither
|
|
474
|
+
provider exposes a size envelope, so none is claimed.
|
|
475
|
+
|
|
476
|
+
## `"recraft/recraft-v4.1": {`
|
|
477
|
+
|
|
478
|
+
---
|
|
479
|
+
|
|
480
|
+
Recraft — V4.1
|
|
481
|
+
-----------------------------------------------------------------
|
|
482
|
+
|
|
483
|
+
Universal: $0.035/img raster on every provider (Vercel AI Gateway feed
|
|
484
|
+
`pricing.image`, OpenRouter endpoint `cost_usd`, fal page payload),
|
|
485
|
+
verified 2026-09-02. Vector styles are $0.08 and a separate route on
|
|
486
|
+
fal/OpenRouter (`.../text-to-vector`, `recraft-v4.1-vector`) — not
|
|
487
|
+
catalogued; `styles: null` here means the raster route only.
|
|
488
|
+
OpenRouter org slug is `recraft`, not `recraft-ai`.
|
|
489
|
+
|
|
490
|
+
## `references: { id: "recraft/recraft-v4.1", max: 1 },`
|
|
491
|
+
|
|
492
|
+
OpenRouter `supported_parameters.input_references` 0–1
|
|
493
|
+
(2026-09-02).
|
|
494
|
+
|
|
495
|
+
## `"recraft/recraft-v3": {`
|
|
496
|
+
|
|
497
|
+
---
|
|
498
|
+
|
|
499
|
+
Recraft — V3
|
|
500
|
+
-----------------------------------------------------------------
|
|
501
|
+
|
|
502
|
+
$0.04/img raster on every provider (re-verified 2026-09-02; Recraft's
|
|
503
|
+
own pricing table agrees).
|
|
504
|
+
|
|
505
|
+
## `pricing: { type: "per_run_flat", usd: 0.04 },`
|
|
506
|
+
|
|
507
|
+
Source: replicate.com/google/lyria-3 — "$0.04 per output audio file"
|
|
508
|
+
|
|
509
|
+
## `pricing: { type: "per_run_flat", usd: 0.08 },`
|
|
510
|
+
|
|
511
|
+
Source: replicate.com/google/lyria-3-pro — "$0.08 per output audio file"
|
|
512
|
+
|
|
513
|
+
## `pricing: {`
|
|
514
|
+
|
|
515
|
+
API: 100 credits when duration is automatic, or 11 credits/s when set.
|
|
516
|
+
|
|
517
|
+
## `pricing: { type: "per_1000_characters", usd: 0.1 },`
|
|
518
|
+
|
|
519
|
+
Source: elevenlabs.io/pricing/api — $0.10 per 1,000 characters.
|
|
520
|
+
|
|
521
|
+
## `"tripo/rig-v2.5": {`
|
|
522
|
+
|
|
523
|
+
Follow the documented compatibility table, not its contradictory biped example.
|
|
524
|
+
|
|
525
|
+
## `"google/veo-3.1": {`
|
|
526
|
+
|
|
527
|
+
---
|
|
528
|
+
|
|
529
|
+
Google — Veo 3.1
|
|
530
|
+
-----------------------------------------------------------------
|
|
531
|
+
|
|
532
|
+
## `vercel: {`
|
|
533
|
+
|
|
534
|
+
Vercel AI Gateway — vercelAiGateway.videoModel(id), image-to-video. The
|
|
535
|
+
Vercel AI Gateway meters both audio modes and sells 4K; the matrix is
|
|
536
|
+
identical to fal's. Verified against Vercel AI Gateway's own
|
|
537
|
+
/v1/models feed (`video_duration_pricing`) on 2026-09-02 — the
|
|
538
|
+
previous card claimed "audio-on only, ≤1080p", which the feed
|
|
539
|
+
contradicts.
|
|
540
|
+
|
|
541
|
+
`silent` is catalogued as provider truth but is not yet
|
|
542
|
+
reachable on Grida's hosted route: the video wire carries no
|
|
543
|
+
audio-mode field, so `generateVideo` prices the card's default
|
|
544
|
+
mode ([`editor/lib/ai/server.ts`](https://github.com/gridaco/grida/blob/main/editor/lib/ai/server.ts) in gridaco/grida). `"4k"` IS reachable — the
|
|
545
|
+
request path maps a 2160 short edge to that label and today
|
|
546
|
+
rejects it for want of a rate.
|
|
547
|
+
https://vercel.com/ai-gateway/models/veo-3.1-generate-001
|
|
548
|
+
|
|
549
|
+
## `fal: {`
|
|
550
|
+
|
|
551
|
+
fal.ai — image-to-video endpoint (capability is keyed into the id;
|
|
552
|
+
t2v is a separate `fal-ai/veo3.1` endpoint, not catalogued). Rate
|
|
553
|
+
matrix is identical to the Vercel AI Gateway's: $0.40/s audio and
|
|
554
|
+
$0.20/s silent at 720p/1080p, $0.60/$0.40 at 4K.
|
|
555
|
+
https://fal.ai/models/fal-ai/veo3.1/image-to-video
|
|
556
|
+
|
|
557
|
+
## `openrouter: {`
|
|
558
|
+
|
|
559
|
+
OpenRouter — async `/api/v1/videos` (job → poll → unsigned url).
|
|
560
|
+
Text-to-video + image-to-video; native audio. "from $0.40/s"
|
|
561
|
+
(verified 2026-06-29, https://openrouter.ai/google/veo-3.1).
|
|
562
|
+
|
|
563
|
+
## `"google/veo-3.1-fast": {`
|
|
564
|
+
|
|
565
|
+
---
|
|
566
|
+
|
|
567
|
+
Google — Veo 3.1 Fast
|
|
568
|
+
-----------------------------------------------------------------
|
|
569
|
+
|
|
570
|
+
Same envelope as Veo 3.1 (16:9/9:16, 4/6/8s, native audio, ≤4K) at
|
|
571
|
+
~2.7x lower cost. Rate matrix is identical on Vercel AI Gateway and fal —
|
|
572
|
+
Vercel AI Gateway feed `video_duration_pricing` and fal's stated per-second
|
|
573
|
+
rates, verified 2026-09-02.
|
|
574
|
+
|
|
575
|
+
## `"google/veo-3.1-lite": {`
|
|
576
|
+
|
|
577
|
+
---
|
|
578
|
+
|
|
579
|
+
Google — Veo 3.1 Lite
|
|
580
|
+
-----------------------------------------------------------------
|
|
581
|
+
|
|
582
|
+
The budget Veo: 720p/1080p only, 4/6/8s, native audio. Rates
|
|
583
|
+
identical on Vercel AI Gateway and fal (verified 2026-09-02). Vercel AI Gateway
|
|
584
|
+
supports both text and image input (FAQ verified 2026-09-07); the fal binding
|
|
585
|
+
below requires an image. Hosted GG's narrower wire is a host concern.
|
|
586
|
+
|
|
587
|
+
## `"alibaba/wan-3.0": {`
|
|
588
|
+
|
|
589
|
+
---
|
|
590
|
+
|
|
591
|
+
Alibaba — Wan 3.0
|
|
592
|
+
-----------------------------------------------------------------
|
|
593
|
+
|
|
594
|
+
Per-second by resolution, audio bundled into the rate (Vercel AI Gateway
|
|
595
|
+
feed has no audio axis; fal exposes an `audio` toggle but bills the
|
|
596
|
+
same). Identical on Vercel AI Gateway and fal, verified 2026-09-02. 2–30s.
|
|
597
|
+
|
|
598
|
+
## `"bytedance/seedance-2.0": {`
|
|
599
|
+
|
|
600
|
+
ByteDance — Seedance 2.0
|
|
601
|
+
-----------------------------------------------------------------
|
|
602
|
+
|
|
603
|
+
## `providers: {`
|
|
604
|
+
|
|
605
|
+
NO Vercel AI Gateway binding, deliberately. Vercel AI Gateway serves
|
|
606
|
+
`bytedance/seedance-2.0` but meters it PER TOKEN
|
|
607
|
+
(`video_token_pricing`: $7.00/MTok at 480p/720p, $7.70 at 1080p,
|
|
608
|
+
$4.00 at 4K; reduced with video input; "minimum token floors based
|
|
609
|
+
on output duration"). `PerSecondPricing` cannot express that, the
|
|
610
|
+
hosted route pre-prices rate×duration, and ByteDance publishes no
|
|
611
|
+
tokens-per-second figure — so there is no honest per-second rate,
|
|
612
|
+
and the per-second Vercel AI Gateway binding this card shipped with was
|
|
613
|
+
invented. Withheld until the hosted path can meter post-flight;
|
|
614
|
+
hosted requests fail with "not available on the hosted provider".
|
|
615
|
+
See gridaco/grida#1019 (Blocker A). Feed checked 2026-09-02.
|
|
616
|
+
|
|
617
|
+
## `fal: {`
|
|
618
|
+
|
|
619
|
+
fal — image-to-video. fal also meters tokens underneath
|
|
620
|
+
($0.014/1K), but states a per-second price for the two
|
|
621
|
+
resolutions it prices: $0.3034/s @720p, $0.682/s @1080p, audio
|
|
622
|
+
bundled. Those are fal's own figures, not a conversion of ours.
|
|
623
|
+
https://fal.ai/models/bytedance/seedance-2.0/image-to-video
|
|
624
|
+
|
|
625
|
+
## `openrouter: {`
|
|
626
|
+
|
|
627
|
+
OpenRouter — async `/api/v1/videos`. Flat $0.06726/s (verified
|
|
628
|
+
2026-06-29, https://openrouter.ai/bytedance/seedance-2.0) — far
|
|
629
|
+
below Vercel AI Gateway's per-resolution rate (the proprietary-pricing-
|
|
630
|
+
diverges finding, #908). No separate silent meter surfaced.
|
|
631
|
+
|
|
632
|
+
## `"bytedance/seedance-2.5": {`
|
|
633
|
+
|
|
634
|
+
---
|
|
635
|
+
|
|
636
|
+
ByteDance — Seedance 2.5
|
|
637
|
+
-----------------------------------------------------------------
|
|
638
|
+
|
|
639
|
+
Newer generation (2026-07-31), NOT a drop-in successor: ~55% more
|
|
640
|
+
per token on Vercel AI Gateway ($10.70/MTok at 480p/720p, $11.70 at 1080p
|
|
641
|
+
vs 2.0's $7.00/$7.70), ~56–71% more per second on fal, and no 4K.
|
|
642
|
+
It buys 4–30s clips, video-editing and extend-video. 2.0 is cheaper
|
|
643
|
+
and serves 4K. Vercel AI Gateway bills tokens, so there is no comparable
|
|
644
|
+
per-second Vercel AI Gateway meter here. fal states per-second prices, audio bundled.
|
|
645
|
+
|
|
646
|
+
## `fal: {`
|
|
647
|
+
|
|
648
|
+
fal's stated rates: $0.2205/s @480p, $0.4730/s @720p, $1.164/s
|
|
649
|
+
@1080p (fal meters $0.0214/1K tokens underneath). 2026-09-02.
|
|
650
|
+
https://fal.ai/models/bytedance/seedance-2.5/image-to-video
|
|
651
|
+
|
|
652
|
+
## `"xai/grok-imagine-video-1.5": {`
|
|
653
|
+
|
|
654
|
+
---
|
|
655
|
+
|
|
656
|
+
SpaceXAI — Grok Imagine Video 1.5
|
|
657
|
+
-----------------------------------------------------------------
|
|
658
|
+
|
|
659
|
+
Image-to-video only (no t2v, per SpaceXAI docs); native lip-synced audio
|
|
660
|
+
bundled into the rate. Per-second by resolution, identical on Vercel AI Gateway
|
|
661
|
+
(no markup) and fal: $0.08/s @480p, $0.14/s @720p, $0.25/s @1080p.
|
|
662
|
+
Both also bill $0.01 per input image, captured separately from the
|
|
663
|
+
output meter.
|
|
664
|
+
|
|
665
|
+
## `vercel: {`
|
|
666
|
+
|
|
667
|
+
Vercel AI Gateway — image-to-video; mirrors SpaceXAI's list price (no markup).
|
|
668
|
+
Vercel AI Gateway namespaces every SpaceXAI model under `spacexai/`, so the
|
|
669
|
+
call id deliberately differs from this card's canonical `xai/` id.
|
|
670
|
+
https://vercel.com/changelog/grok-imagine-video-1-5-on-ai-gateway
|
|
671
|
+
|
|
672
|
+
## `fal: {`
|
|
673
|
+
|
|
674
|
+
fal.ai — image-to-video endpoint; same per-second rate.
|
|
675
|
+
https://fal.ai/models/xai/grok-imagine-video/v1.5/image-to-video
|
|
676
|
+
|
|
677
|
+
## `pricing: { type: "per_token_input", input: 0.15 },`
|
|
678
|
+
|
|
679
|
+
VERIFY before prod: models.dev lists no per-1M price for
|
|
680
|
+
gemini-embedding-2 yet; using the gemini-embedding-001 family
|
|
681
|
+
input rate ($0.15 / 1M) as a conservative stand-in.
|